Hexagon Accelerates AI Model Training with SageMaker HyperPod
Hexagon, a global leader in measurement technologies, develops specialized AI models to process vast amounts of 3D point cloud data for critical industries. These models perform complex tasks like noise removal and land classification, demanding robust and scalable training infrastructure. To accelerate AI innovation and reduce model development time, Hexagon partnered with AWS to implement Amazon SageMaker HyperPod.
SageMaker HyperPod is an ML infrastructure designed for large-scale, distributed training of foundation models. Its key features include a resilient architecture with built-in self-healing capabilities and automated job resumption, enabling training runs to continue uninterrupted for weeks or months. It offers scalable infrastructure through single-spine node topology and pre-configured Elastic Fabric Adapter (EFA), ensuring optimal inter-node communication and flexible compute capacity, ideal for multi-node GPU training. HyperPod supports versatile deployment, compatible with various generative AI software stacks and leading Amazon EC2 instances like P6-B200 and P6e-GB200, accelerated by NVIDIA Blackwell GPUs. Operational efficiency is enhanced via intelligent task governance, pre-configured Deep Learning AMIs, and quick start training recipes.
Hexagon’s implementation utilized an integrated data pipeline, storing training data in Amazon S3 and leveraging Amazon FSx for Lustre for high-performance streaming directly to GPU nodes at multi-GB/s. The HyperPod cluster managed compute with built-in health checks and automated instance management, while SageMaker Training Plans provided predictable GPU capacity reservation. For MLOps, a one-click observability solution published metrics to Prometheus and Grafana, offering per-GPU level monitoring. MLflow on SageMaker AI tracked experiments with minimal code changes.
Benefits for Hexagon were significant: integration and first training deployment were achieved within hours. Training time for a specific model was dramatically reduced by 95%, from 80 days on-premises to approximately 4 days on AWS using 6x ml.p5.48xlarge instances, each with eight NVIDIA H100 GPUs and EFA. This performance enhancement also enabled larger batch sizes, leading to higher accuracy scores for their specialized AI models. SageMaker HyperPod empowers Hexagon to accelerate innovation and develop next-generation AI products faster.
Hexagon’s implementation of SageMaker HyperPod demonstrates how advanced infrastructure can streamline ai automation training workflows for enterprise applications.
While Hexagon’s SageMaker HyperPod implementation focuses on industrial AI applications, similar infrastructure could theoretically support chatgpt automation training workflows at scale.
